Update search_utils.py

search_utils.py

@@ -1,44 +1,225 @@
 import numpy as np
+import pandas as pd
 import faiss
 import zipfile
 import logging
 from pathlib import Path
-from sentence_transformers import SentenceTransformer
-import concurrent.futures
+from sentence_transformers import SentenceTransformer, util
+import streamlit as st
+import time
 import os
+from urllib.parse import quote
 import requests
+import shutil
+import concurrent.futures
+# Optional: Uncomment if you want to use lru_cache for instance methods
 from functools import lru_cache
-from typing import List, Dict
 
 # Configure logging
-logging.basicConfig(level=logging.WARNING,
-                    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
-logger = logging.getLogger("OptimizedSearch")
+logging.basicConfig(
+    level=logging.INFO,
+    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s',
+    handlers=[logging.StreamHandler()]
+)
+logger = logging.getLogger("MetadataManager")
 
-class OptimizedMetadataManager:
+class MetadataManager:
     def __init__(self):
-        self._init_metadata()
+        self.cache_dir = Path("unzipped_cache")
+        self.shard_dir = self.cache_dir / "metadata_shards"
+        self.shard_map = {}
+        self.loaded_shards = {}
+        self.total_docs = 0
+        self.api_cache = {}
+        
+        logger.info("Initializing MetadataManager")
+        self._ensure_directories()
+        self._unzip_if_needed()
+        self._build_shard_map()
         self._init_url_resolver()
-
-    def _init_metadata(self):
-        """Memory-mapped metadata loading"""
-        self.metadata_dir = Path("unzipped_cache/metadata_shards")
-        self.metadata = {}
+        logger.info(f"Total documents indexed: {self.total_docs}")
+        logger.info(f"Total shards found: {len(self.shard_map)}")
+    
+    def _ensure_directories(self):
+        """Create necessary directories if they don't exist."""
+        self.cache_dir.mkdir(parents=True, exist_ok=True)
+        self.shard_dir.mkdir(parents=True, exist_ok=True)
+    
+    def _unzip_if_needed(self):
+        """Extract the ZIP archive if no parquet files are found."""
+        zip_path = Path("metadata_shards.zip")
+        if not any(self.shard_dir.rglob("*.parquet")):
+            logger.info("No parquet files found, checking for zip archive")
+            if not zip_path.exists():
+                raise FileNotFoundError(f"Metadata ZIP file not found at {zip_path}")
+            logger.info(f"Extracting {zip_path} to {self.shard_dir}")
+            try:
+                with zipfile.ZipFile(zip_path, 'r') as zip_ref:
+                    zip_root = self._get_zip_root(zip_ref)
+                    zip_ref.extractall(self.shard_dir)
+                    if zip_root:
+                        nested_dir = self.shard_dir / zip_root
+                        if nested_dir.exists():
+                            self._flatten_directory(nested_dir, self.shard_dir)
+                            nested_dir.rmdir()
+                    parquet_files = list(self.shard_dir.rglob("*.parquet"))
+                    if not parquet_files:
+                        raise RuntimeError("Extraction completed but no parquet files found")
+                    logger.info(f"Found {len(parquet_files)} parquet files after extraction")
+            except Exception as e:
+                logger.error(f"Failed to extract zip file: {str(e)}")
+                self._clean_failed_extraction()
+                raise
+    
+    def _get_zip_root(self, zip_ref):
+        """Identify the common root directory within the ZIP file."""
+        try:
+            first_file = zip_ref.namelist()[0]
+            if '/' in first_file:
+                return first_file.split('/')[0]
+            return ""
+        except Exception as e:
+            logger.warning(f"Error detecting zip root: {str(e)}")
+            return ""
+    
+    def _flatten_directory(self, src_dir, dest_dir):
+        """Move files from a nested directory up to the destination."""
+        for item in src_dir.iterdir():
+            if item.is_dir():
+                self._flatten_directory(item, dest_dir)
+                item.rmdir()
+            else:
+                target = dest_dir / item.name
+                if target.exists():
+                    target.unlink()
+                item.rename(target)
+    
+    def _clean_failed_extraction(self):
+        """Clean up files from a failed extraction attempt."""
+        logger.info("Cleaning up failed extraction")
+        for item in self.shard_dir.iterdir():
+            if item.is_dir():
+                shutil.rmtree(item)
+            else:
+                item.unlink()
+    
+    def _build_shard_map(self):
+        """Build a map from global index ranges to shard filenames."""
+        logger.info("Building shard map from parquet files")
+        parquet_files = list(self.shard_dir.glob("*.parquet"))
+        if not parquet_files:
+            raise FileNotFoundError("No parquet files found after extraction")
+        parquet_files = sorted(parquet_files, key=lambda x: int(x.stem.split("_")[1]))
+        expected_start = 0
+        for f in parquet_files:
+            try:
+                parts = f.stem.split("_")
+                if len(parts) != 3:
+                    raise ValueError("Invalid filename format")
+                start = int(parts[1])
+                end = int(parts[2])
+                if start != expected_start:
+                    raise ValueError(f"Non-contiguous shard start: expected {expected_start}, got {start}")
+                if end <= start:
+                    raise ValueError(f"Invalid shard range: {start}-{end}")
+                self.shard_map[(start, end)] = f.name
+                self.total_docs = end + 1
+                expected_start = end + 1
+                logger.debug(f"Mapped shard {f.name}: indices {start}-{end}")
+            except Exception as e:
+                logger.error(f"Error processing shard {f.name}: {str(e)}")
+                raise RuntimeError("Invalid shard structure") from e
+        logger.info(f"Validated {len(self.shard_map)} continuous shards")
+        logger.info(f"Total document count: {self.total_docs}")
+        sorted_ranges = sorted(self.shard_map.keys())
+        for i in range(1, len(sorted_ranges)):
+            prev_end = sorted_ranges[i-1][1]
+            curr_start = sorted_ranges[i][0]
+            if curr_start != prev_end + 1:
+                logger.warning(f"Gap or overlap detected between shards: {prev_end} to {curr_start}")
+    
+    def _process_shard(self, shard, local_indices):
+        """Load a shard (if not already loaded) and retrieve the specified rows."""
+        try:
+            if shard not in self.loaded_shards:
+                shard_path = self.shard_dir / shard
+                if not shard_path.exists():
+                    logger.error(f"Shard file not found: {shard_path}")
+                    return pd.DataFrame(columns=["title", "summary", "similarity"])
+                file_size_mb = os.path.getsize(shard_path) / (1024 * 1024)
+                logger.info(f"Loading shard file: {shard} (size: {file_size_mb:.2f} MB)")
+                try:
+                    self.loaded_shards[shard] = pd.read_parquet(shard_path, columns=["title", "summary"])
+                    logger.info(f"Loaded shard {shard} with {len(self.loaded_shards[shard])} rows")
+                except Exception as e:
+                    logger.error(f"Failed to read parquet file {shard}: {str(e)}")
+                    try:
+                        schema = pd.read_parquet(shard_path, engine='pyarrow').dtypes
+                        logger.info(f"Parquet schema: {schema}")
+                    except Exception:
+                        pass
+                    return pd.DataFrame(columns=["title", "summary", "similarity"])
+            df = self.loaded_shards[shard]
+            df_len = len(df)
+            valid_local_indices = [idx for idx in local_indices if 0 <= idx < df_len]
+            if len(valid_local_indices) != len(local_indices):
+                logger.warning(f"Filtered {len(local_indices) - len(valid_local_indices)} out-of-bounds indices in shard {shard}")
+            if valid_local_indices:
+                chunk = df.iloc[valid_local_indices]
+                logger.info(f"Retrieved {len(chunk)} records from shard {shard}")
+                return chunk
+        except Exception as e:
+            logger.error(f"Error processing shard {shard}: {str(e)}", exc_info=True)
+        return pd.DataFrame(columns=["title", "summary", "similarity"])
+    
+    def get_metadata(self, global_indices):
+        """Retrieve metadata for a batch of global indices using parallel shard processing."""
+        if isinstance(global_indices, np.ndarray) and global_indices.size == 0:
+            logger.warning("Empty indices array passed to get_metadata")
+            return pd.DataFrame(columns=["title", "summary", "similarity"])
         
-        # Preload all metadata into memory
-        for parquet_file in self.metadata_dir.glob("*.parquet"):
-            df = pd.read_parquet(parquet_file, columns=["title", "summary"])
-            self.metadata.update(df.to_dict(orient="index"))
-            
-        self.total_docs = len(self.metadata)
-        logger.info(f"Loaded {self.total_docs} metadata entries into memory")
-
-    def get_metadata_batch(self, indices: np.ndarray) -> List[Dict]:
-        """Batch retrieval of metadata"""
-        return [self.metadata.get(idx, {"title": "", "summary": ""}) for idx in indices]
-
+        indices_list = global_indices.tolist() if isinstance(global_indices, np.ndarray) else global_indices
+        logger.info(f"Retrieving metadata for {len(indices_list)} indices")
+        valid_indices = [idx for idx in indices_list if 0 <= idx < self.total_docs]
+        invalid_count = len(indices_list) - len(valid_indices)
+        if invalid_count > 0:
+            logger.warning(f"Filtered out {invalid_count} invalid indices")
+        if not valid_indices:
+            logger.warning("No valid indices remain after filtering")
+            return pd.DataFrame(columns=["title", "summary", "similarity"])
+        
+        # Group indices by shard
+        shard_groups = {}
+        for idx in valid_indices:
+            found = False
+            for (start, end), shard in self.shard_map.items():
+                if start <= idx <= end:
+                    shard_groups.setdefault(shard, []).append(idx - start)
+                    found = True
+                    break
+            if not found:
+                logger.warning(f"Index {idx} not found in any shard range")
+        
+        # Process shards concurrently
+        results = []
+        with concurrent.futures.ThreadPoolExecutor() as executor:
+            futures = [executor.submit(self._process_shard, shard, local_indices)
+                       for shard, local_indices in shard_groups.items()]
+            for future in concurrent.futures.as_completed(futures):
+                df_chunk = future.result()
+                if not df_chunk.empty:
+                    results.append(df_chunk)
+        
+        if results:
+            combined = pd.concat(results).reset_index(drop=True)
+            logger.info(f"Combined metadata: {len(combined)} records from {len(results)} shards")
+            return combined
+        else:
+            logger.warning("No metadata records retrieved")
+            return pd.DataFrame(columns=["title", "summary", "similarity"])
+    
     def _init_url_resolver(self):
-        """Initialize API session and cache"""
+        """Initialize API session and cache."""
         self.session = requests.Session()
         adapter = requests.adapters.HTTPAdapter(
             pool_connections=10,
@@ -47,43 +228,45 @@ class OptimizedMetadataManager:
         )
         self.session.mount("https://", adapter)
 
-    @lru_cache(maxsize=10_000)
     def resolve_url(self, title: str) -> str:
-        """Optimized URL resolution with fail-fast"""
-        try:
-            # Try arXiv first
-            arxiv_url = self._get_arxiv_url(title)
-            if arxiv_url: return arxiv_url
-            
-            # Fallback to Semantic Scholar
-            semantic_url = self._get_semantic_url(title)
-            if semantic_url: return semantic_url
-            
-        except Exception as e:
-            logger.warning(f"URL resolution failed: {str(e)}")
-            
-        return f"https://scholar.google.com/scholar?q={quote(title)}"
-
+        """Optimized URL resolution with fail-fast."""
+        if title in self.api_cache:
+            return self.api_cache[title]
+        
+        links = {}
+        arxiv_url = self._get_arxiv_url(title)
+        if arxiv_url:
+            links["arxiv"] = arxiv_url
+        semantic_url = self._get_semantic_url(title)
+        if semantic_url:
+            links["semantic"] = semantic_url
+        scholar_url = f"https://scholar.google.com/scholar?q={quote(title)}"
+        links["google"] = scholar_url
+        
+        self.api_cache[title] = links
+        return links
+    
     def _get_arxiv_url(self, title: str) -> str:
-        """Fast arXiv lookup with timeout"""
+        """Fast arXiv lookup with timeout."""
         with self.session.get(
             "http://export.arxiv.org/api/query",
-            params={"search_query": f'ti:"{title}"', "max_results": 1},
+            params={"search_query": f'ti:"{title}"', "max_results": 1, "sortBy": "relevance"},
             timeout=2
         ) as response:
             if response.ok:
                 return self._parse_arxiv_response(response.text)
         return ""
-
+    
     def _parse_arxiv_response(self, xml: str) -> str:
-        """Fast XML parsing using string operations"""
-        if "<entry>" not in xml: return ""
+        """Fast XML parsing using string operations."""
+        if "<entry>" not in xml:
+            return ""
         start = xml.find("<id>") + 4
         end = xml.find("</id>", start)
         return xml[start:end].replace("http:", "https:") if start > 3 else ""
-
+    
     def _get_semantic_url(self, title: str) -> str:
-        """Batch-friendly Semantic Scholar lookup"""
+        """Batch-friendly Semantic Scholar lookup."""
         with self.session.get(
             "https://api.semanticscholar.org/graph/v1/paper/search",
             params={"query": title[:200], "limit": 1},
@@ -94,61 +277,190 @@ class OptimizedMetadataManager:
                 if data.get("data"):
                     return data["data"][0].get("url", "")
         return ""
+    
+    def _format_source_links(self, links):
+        """Generate an HTML snippet for the available source links."""
+        html_parts = []
+        if "arxiv" in links:
+            html_parts.append(f"<a class='source-link' href='{links['arxiv']}' target='_blank' rel='noopener noreferrer'> 📜 arXiv</a>")
+        if "semantic" in links:
+            html_parts.append(f"<a class='source-link' href='{links['semantic']}' target='_blank' rel='noopener noreferrer'> 🌐 Semantic Scholar</a>")
+        if "google" in links:
+            html_parts.append(f"<a class='source-link' href='{links['google']}' target='_blank' rel='noopener noreferrer'> 🔍 Google Scholar</a>")
+        return " | ".join(html_parts)
 
-class OptimizedSemanticSearch:
-    def __init__(self):
-        self.model = SentenceTransformer('all-MiniLM-L6-v2')
-        self._load_faiss_indexes()
-        self.metadata_mgr = OptimizedMetadataManager()
-        
-    def _load_faiss_indexes(self):
-        """Load indexes with memory mapping"""
-        self.index = faiss.read_index("combined_index.faiss", faiss.IO_FLAG_MMAP | faiss.IO_FLAG_READ_ONLY)
-        logger.info(f"Loaded FAISS index with {self.index.ntotal} vectors")
 
-    def search(self, query: str, top_k: int = 5) -> List[Dict]:
-        """Optimized search pipeline"""
-        # Batch encode query
-        query_embedding = self.model.encode([query], convert_to_numpy=True)
-        
-        # FAISS search
-        distances, indices = self.index.search(query_embedding, top_k*2)  # Search extra for dedup
-        
-        # Batch metadata retrieval
-        results = self.metadata_mgr.get_metadata_batch(indices[0])
-        
-        # Process results
-        return self._process_results(results, distances[0], top_k)
-
-    def _process_results(self, results: List[Dict], distances: np.ndarray, top_k: int) -> List[Dict]:
-        """Parallel result processing"""
+class SemanticSearch:
+    def __init__(self):
+        self.shard_dir = Path("compressed_shards")
+        self.model = None
+        self.index_shards = []
+        self.metadata_mgr = MetadataManager()
+        self.shard_sizes = []
+        self.cumulative_offsets = None
+        self.logger = logging.getLogger("SemanticSearch")
+        self.logger.info("Initializing SemanticSearch")
+    
+    @st.cache_resource
+    def load_model(_self):
+        return SentenceTransformer('all-MiniLM-L6-v2')
+    
+    def initialize_system(self):
+        self.logger.info("Loading sentence transformer model")
+        start_time = time.time()
+        self.model = self.load_model()
+        self.logger.info(f"Model loaded in {time.time() - start_time:.2f} seconds")
+        self.logger.info("Loading FAISS indices")
+        self._load_faiss_shards()
+    
+    def _load_faiss_shards(self):
+        """Load FAISS shards concurrently and precompute cumulative offsets for global indexing."""
+        self.logger.info(f"Searching for index files in {self.shard_dir}")
+        if not self.shard_dir.exists():
+            self.logger.error(f"Shard directory not found: {self.shard_dir}")
+            return
+        index_files = sorted(self.shard_dir.glob("*.index"))
+        self.logger.info(f"Found {len(index_files)} index files")
+        self.index_shards = []
+        self.shard_sizes = []
         with concurrent.futures.ThreadPoolExecutor() as executor:
-            # Parallel URL resolution
-            futures = {
-                executor.submit(
-                    self.metadata_mgr.resolve_url, 
-                    res["title"]
-                ): idx for idx, res in enumerate(results)
+            future_to_file = {
+                executor.submit(self._load_single_index, shard_path): shard_path
+                for shard_path in index_files
             }
-            
-            # Update results as URLs resolve
-            for future in concurrent.futures.as_completed(futures):
-                idx = futures[future]
+            for future in concurrent.futures.as_completed(future_to_file):
+                shard_path = future_to_file[future]
                 try:
-                    results[idx]["source"] = future.result()
+                    index, size = future.result()
+                    if index is not None:
+                        self.index_shards.append(index)
+                        self.shard_sizes.append(size)
+                        self.logger.info(f"Loaded index {shard_path.name} with {size} vectors")
                 except Exception as e:
-                    results[idx]["source"] = ""
-
-        # Add similarity scores
-        for idx, dist in enumerate(distances[:len(results)]):
-            results[idx]["similarity"] = 1 - (dist / 2)
-
-        # Deduplicate and sort
-        seen = set()
-        final_results = []
-        for res in sorted(results, key=lambda x: x["similarity"], reverse=True):
-            if res["title"] not in seen and len(final_results) < top_k:
-                seen.add(res["title"])
-                final_results.append(res)
+                    self.logger.error(f"Error loading index {shard_path}: {str(e)}")
+        total_vectors = sum(self.shard_sizes)
+        self.logger.info(f"Total loaded vectors: {total_vectors} across {len(self.index_shards)} shards")
+        self.cumulative_offsets = np.cumsum([0] + self.shard_sizes)
+    
+    def _load_single_index(self, shard_path):
+        """Load a single FAISS index shard."""
+        self.logger.info(f"Loading index: {shard_path}")
+        start_time = time.time()
+        file_size_mb = os.path.getsize(shard_path) / (1024 * 1024)
+        self.logger.info(f"Index file size: {file_size_mb:.2f} MB")
+        index = faiss.read_index(str(shard_path))
+        size = index.ntotal
+        self.logger.info(f"Index loaded in {time.time() - start_time:.2f} seconds")
+        return index, size
+    
+    def _global_index(self, shard_idx, local_idx):
+        """Convert a local index (within a shard) to a global index using precomputed offsets."""
+        return int(self.cumulative_offsets[shard_idx] + local_idx)
+    
+    def search(self, query, top_k=5):
+        """Search for a query using parallel FAISS shard search."""
+        self.logger.info(f"Searching for query: '{query}' (top_k={top_k})")
+        start_time = time.time()
+        if not query:
+            self.logger.warning("Empty query provided")
+            return pd.DataFrame()
+        if not self.index_shards:
+            self.logger.error("No index shards loaded")
+            return pd.DataFrame()
+        try:
+            self.logger.info("Encoding query")
+            query_embedding = self.model.encode([query], convert_to_numpy=True)
+            self.logger.debug(f"Query encoded to shape {query_embedding.shape}")
+        except Exception as e:
+            self.logger.error(f"Query encoding failed: {str(e)}")
+            return pd.DataFrame()
         
-        return final_results
+        all_distances = []
+        all_global_indices = []
+        # Run shard searches in parallel
+        with concurrent.futures.ThreadPoolExecutor() as executor:
+            futures = {
+                executor.submit(self._search_shard, shard_idx, index, query_embedding, top_k): shard_idx
+                for shard_idx, index in enumerate(self.index_shards)
+            }
+            for future in concurrent.futures.as_completed(futures):
+                result = future.result()
+                if result is not None:
+                    distances_part, global_indices_part = result
+                    all_distances.extend(distances_part)
+                    all_global_indices.extend(global_indices_part)
+        self.logger.info(f"Search found {len(all_global_indices)} results across all shards")
+        results = self._process_results(np.array(all_distances), np.array(all_global_indices), top_k)
+        self.logger.info(f"Search completed in {time.time() - start_time:.2f} seconds with {len(results)} final results")
+        return results
+    
+    def _search_shard(self, shard_idx, index, query_embedding, top_k):
+        """Search a single FAISS shard for the query embedding."""
+        if index.ntotal == 0:
+            self.logger.warning(f"Skipping empty shard {shard_idx}")
+            return None
+        try:
+            shard_start = time.time()
+            distances, indices = index.search(query_embedding, top_k)
+            valid_mask = (indices[0] >= 0) & (indices[0] < index.ntotal)
+            valid_indices = indices[0][valid_mask].tolist()
+            valid_distances = distances[0][valid_mask].tolist()
+            if len(valid_indices) != top_k:
+                self.logger.debug(f"Shard {shard_idx}: Found {len(valid_indices)} valid results out of {top_k}")
+            global_indices = [self._global_index(shard_idx, idx) for idx in valid_indices]
+            self.logger.debug(f"Shard {shard_idx} search completed in {time.time() - shard_start:.3f}s")
+            return valid_distances, global_indices
+        except Exception as e:
+            self.logger.error(f"Search failed in shard {shard_idx}: {str(e)}")
+            return None
+    
+    def _process_results(self, distances, global_indices, top_k):
+        """Process raw search results: retrieve metadata, calculate similarity, and deduplicate."""
+        process_start = time.time()
+        if global_indices.size == 0 or distances.size == 0:
+            self.logger.warning("No search results to process")
+            return pd.DataFrame(columns=["title", "summary", "source", "similarity"])
+        try:
+            self.logger.info(f"Retrieving metadata for {len(global_indices)} indices")
+            metadata_start = time.time()
+            results = self.metadata_mgr.get_metadata(global_indices)
+            self.logger.info(f"Metadata retrieved in {time.time() - metadata_start:.2f}s, got {len(results)} records")
+            if len(results) == 0:
+                self.logger.warning("No metadata found for indices")
+                return pd.DataFrame(columns=["title", "summary", "source", "similarity"])
+            if len(results) != len(distances):
+                self.logger.warning(f"Mismatch between distances ({len(distances)}) and results ({len(results)})")
+                if len(results) < len(distances):
+                    distances = distances[:len(results)]
+                else:
+                    distances = np.pad(distances, (0, len(results) - len(distances)), 'constant', constant_values=1.0)
+            self.logger.debug("Calculating similarity scores")
+            results['similarity'] = 1 - (distances / 2)
+            if not results.empty:
+                self.logger.debug(f"Similarity stats: min={results['similarity'].min():.3f}, " +
+                                  f"max={results['similarity'].max():.3f}, " +
+                                  f"mean={results['similarity'].mean():.3f}")
+            results['source'] = results['title'].apply(
+                lambda title: self._format_source_links(self.metadata_mgr._resolve_paper_url(title))
+            )
+            pre_dedup = len(results)
+            results = results.drop_duplicates(subset=["title", "source"]).sort_values("similarity", ascending=False).head(top_k)
+            post_dedup = len(results)
+            if pre_dedup > post_dedup:
+                self.logger.info(f"Removed {pre_dedup - post_dedup} duplicate results")
+            self.logger.info(f"Results processed in {time.time() - process_start:.2f}s, returning {len(results)} items")
+            return results.reset_index(drop=True)
+        except Exception as e:
+            self.logger.error(f"Result processing failed: {str(e)}", exc_info=True)
+            return pd.DataFrame(columns=["title", "summary", "source", "similarity"])
+    
+    def _format_source_links(self, links):
+        """Generate an HTML snippet for the available source links."""
+        html_parts = []
+        if "arxiv" in links:
+            html_parts.append(f"<a class='source-link' href='{links['arxiv']}' target='_blank' rel='noopener noreferrer'> 📜 arXiv</a>")
+        if "semantic" in links:
+            html_parts.append(f"<a class='source-link' href='{links['semantic']}' target='_blank' rel='noopener noreferrer'> 🌐 Semantic Scholar</a>")
+        if "google" in links:
+            html_parts.append(f"<a class='source-link' href='{links['google']}' target='_blank' rel='noopener noreferrer'> 🔍 Google Scholar</a>")
+        return " | ".join(html_parts)